Add ArcFace loss.

[innat]

System information.

TensorFlow version (you are using): 2.8 Are you willing to contribute it (Yes/No) : No (but willing to provide starter if required.)

Describe the feature and the current behavior/state.

ArcFace, or Additive Angular Margin Loss, is a loss function used mainly in face recognition tasks. It is used to obtain highly discriminative features. This loss function can be used in both vision task and nlp task. So, I think its general for keras_cv and keras_nlp.

Paper: https://arxiv.org/pdf/1801.07698.pdf, TL,DR, https://paperswithcode.com/method/arcface Citation: 2939 (current)

Will this change the current api? How?

tf.keras.losses.SparseArcFace
tf.keras.losses.CategoricalArcFace

Who will benefit from this feature?

keras practitioners.

Contributing

• Briefly describe your candidate solution(if contributing):
  • https://github.com/4uiiurz1/keras-arcface

[fchollet]

ArcFace, or Additive Angular Margin Loss, is a loss function used mainly in face recognition tasks

What are some NLP use cases? It seems like it might be a good fit for KerasCV specifically.

[innat]

Sorry I miswrote, it should be only CV task. Could you please move this issue to KerasCV?

[ayulockin]

Hey, @LukeWood @sayakpaul I think this would be a great feature addition to KerasCV. It's especially popular in Kaggle competitions for reference.

Please assign it to me.

[sayakpaul]

I concur. Especially for retrieval tasks, ArcFace loss is greatly beneficial.

[LukeWood]

interesting, can this be used with the existing models we have? i.e. if we add this will we have the tools we need to support facial recognition?

[sayakpaul]

Should be compatible with existing models.

[LukeWood]

So theoretically, we include this, and use this loss to create a facial recognition tutorial with the other existing components?

[sayakpaul]

Shoul be the case. We have a tutorial on few-shot learning and we coul definitely take some components from that too.

[LukeWood]

Shoul be the case. We have a tutorial on few-shot learning and we coul definitely take some components from that too.

ok, sounds good. Then I think this is a great component. Especially if we can include a tutorial on facial recognition.

[sayakpaul]

@ayulockin take it for the grab if you're up.

[LukeWood]

When we do include this I do think it’s important to have a tutorial showing off how to use it, so let’s keep that in mind! Doesn’t have to be anything crazy, just some basics!

[ayulockin]

This loss is compatible with existing models.

I can take this up. Will start working on it soon. Also the idea of tutorial is great for this loss.

[yashchks87]

Would love to work on this if no one has picked up still.

[alexandrekj]

Any updates on this?

[jbischof]

It looks like this fell through the cracks @alexandrekj. @yashchks87 or @alexandrekj, are either of you interested in working on this?

[yashchks87]

I can take on it. @jbischof